AI model licenses explained: can you use an open model commercially?
A plain-English guide to open AI licenses: Apache 2.0, MIT, community licenses with revenue caps, research-only terms and gated downloads, plus a checklist to run before you ship.
Short answer. Check three things before you use an open model for work: whether commercial use is allowed, which conditions come with it (a revenue limit, a credit line, rules for hosting businesses) and what the license says about the outputs. Permissive licenses such as Apache 2.0 and MIT ask for very little. Custom “community” licenses usually add conditions. Research-only licenses forbid products until you agree a separate license. This is a guide, not legal advice.
The three kinds of license on this site
Every model page and the licenses page label the license as one of three types.
| Type | What it means | Examples |
|---|---|---|
| Permissive | Commercial use, changes and redistribution with very few conditions | Apache 2.0, MIT, OpenMDW-1.1, CC BY 4.0 |
| Conditional | Commercial use is allowed, but with conditions such as revenue limits, credit lines or rules for hosting businesses | FLUX Kommunity, Kimi K3, Mistral Open-Weight, Muse Model License, Upstage Solar, BigScience OpenRAIL-M |
| Non-commercial | Research and evaluation only, and commercial use needs a separate license | Qwen Research License |
The models you can use commercially page sorts the whole catalog this way.
The permissive licenses in two minutes
- Apache 2.0. You may use, change, sell and redistribute the model. Keep the license text and copyright notices, and mark files you changed. It includes a patent license, which ends if you sue over a patent on the work.
- MIT. Even shorter: keep the copyright notice and the license text. It does not spell out a patent grant, so some large companies prefer Apache 2.0.
- OpenMDW-1.1. A newer permissive license written for machine-learning models by the Linux Foundation and the PyTorch Foundation. It covers the model, its data, documentation and software, and it puts no restrictions on what you do with outputs. The Open Source Initiative has not approved it.
- CC BY 4.0. A Creative Commons license written for creative and data works rather than software. Some speech releases, such as NVIDIA’s Parakeet TDT 0.6B v3, use it for model weights. It allows commercial use as long as you credit the author.
Conditions you will meet in custom licenses
| Condition | What it asks | Examples from our catalog |
|---|---|---|
| Revenue cap | Free commercial use only while your company, with its affiliates, earns below a yearly limit | FLUX Kommunity (US$5 million), LFM Open 1.0 (US$10 million), LTX-2.x Community (US$10 million), ERNIE 5.5 Flash Base (US$50 million), Chandra OCR 2 (US$2 million) |
| Hosting agreement | A business that sells hosted access to the model needs a separate agreement | Kimi K3 and the Muse Model License (above US$20 million of revenue), Qwen3.8-Max (above US$50 million), Qwen Community 1.0 and Mistral Open-Weight (any size) |
| Credit line | You must show the model’s name | Muse Spark Open (“Built with Muse Spark Open”, for everyone), MiniMax M3.1-Flash (for commercial use), ERNIE 5.5 Flash Base (“Powered by ERNIE”), CC BY 4.0 (credit the author), Kimi K3 and Qwen3.8-Max (only for very large products) |
| Naming rule | A model you derive and share must carry a required name | Upstage Solar License: the name must start with “Solar” and show “Built with Solar” |
| Territory limit | The license is not valid in some regions | ERNIE 5.5 Flash Base and Tencent Hunyuan Community (used by HunyuanOCR): not valid in the European Union, the United Kingdom and South Korea |
| Usage policy | A list of banned uses, such as unlawful content or military use | FLUX Kommunity, LTX-2.x Community, MiniMax Community, ERNIE 5.5 Flash Base, Tencent Hunyuan Community |
| Output rules | Filtering or review of outputs, AI disclosures or keeping watermarks | FLUX Kommunity, LTX-2.x Community |
| Security review | Extra review for the very largest hosting businesses | GLM-5.3 (model-as-a-service above US$10 billion of revenue) |
Read the thresholds carefully. They are measured in different ways: yearly or monthly, for your company or for the whole group of affiliated companies, and in revenue or in users.
OpenRAIL licenses are a special case. They allow commercial use but attach a list of banned uses to every copy and derivative. Supertonic 3 uses BigScience OpenRAIL-M, and Chandra OCR 2 uses Datalab’s modified version, which adds a revenue limit.
Derivatives, fine-tunes and quantized copies
The license travels with the weights. A GGUF or other quantized copy of a model is still covered by the original license. Fine-tunes and distilled models usually count as derivatives and stay under the same terms: the LTX-2.x license says so explicitly, and the LFM Open license lets you redistribute derivative works under the same license. If you plan to publish your own fine-tune, check what the license requires you to include, such as a copy of the license, a notice or a credit line. Fine-tuning for beginners covers the technical side.
Gated downloads are not licenses
Some repositories on Hugging Face are gated: you must log in and accept terms, or request access, before you can download the files. Muse Spark Open, FLUX 3 Image Klein, LTX-2.5 and Cohere Transcribe work this way. Gating is an access step and does not change the license, although you do need an access token for automated downloads.
When labels and files disagree
A license label is only as reliable as the repository behind it. Four real examples from our catalog:
- The Hugging Face page for Whisper large-v3 says Apache 2.0, while the GitHub repository and the large-v3-turbo page say MIT. Both are permissive, but keep the license text for the checkpoint you actually use.
- The MiMo-V2.6 repositories are tagged MIT on Hugging Face, and Xiaomi’s announcement calls the release fully open source. As of 8 November 2026, none of the repositories contains a LICENSE file, so anyone shipping a product should ask Xiaomi to confirm the terms in writing.
- Developers change terms between releases. Black Forest Labs’ earlier FLUX.2 klein 4B is listed under Apache 2.0, and the newer FLUX 3 Image Klein uses the FLUX Kommunity license. Earlier Qwen image models used Apache 2.0 too, and Qwen-Image-2.1 is research-only. Never assume that a new version has the terms of the old one.
- Kyutai’s Pocket TTS weights are listed as CC BY 4.0 on Hugging Face, behind a prohibited-use gate, while the launch post says the model was released under the MIT license. Confirm the terms of the repository you download from.
A checklist before you ship
- Open the LICENSE file in the model’s own repository, not only the label on a listing.
- Confirm commercial use is allowed for your kind of business and size, counting your affiliates.
- Check where you may use it. Some licenses are not valid in the European Union, the United Kingdom or South Korea.
- Check hosting and resale. If you sell API access, fine-tuning or a hosted endpoint, look for model-as-a-service terms.
- Note the credit, naming and notice duties, and put them where the license says: the interface, documentation or website.
- Read the rules on outputs: filtering, AI disclosures, watermarks, and any ban on using outputs to train other models.
- Check the banned-use list for your field, such as military, biometric or regulated uses.
- Keep a copy of the license text and the repository state on the day you downloaded.
- Ask when unclear. For a product with real money behind it, ask the developer, or a lawyer, to confirm.
For a map of the options by license type, see open models you can use commercially, and to understand how licenses relate to the idea of open source, read open weights vs open source.
Frequently asked questions
- Can I use an open-weights model commercially?
- It depends on the license. Apache 2.0, MIT and OpenMDW allow it with light conditions. Custom community licenses often allow it below a revenue limit or with a credit line. Research-only licenses, such as the Qwen Research License, need a separate commercial license.
- What does model as a service mean in a license?
- It means a business that sells other people access to the model, for example through an API, a hosted endpoint or fine-tuning. Licenses such as Kimi K3, Mistral Open-Weight, Muse Model and the Qwen licenses ask such businesses for a separate agreement, with different thresholds. Building the model into a feature of your own product, or simply relaying requests to a model hosted by someone else, is generally not model as a service.
- Does a revenue limit include my affiliates?
- Often yes. The FLUX Kommunity, LFM Open and LTX-2.x licenses measure revenue across the whole corporate group, so a small subsidiary of a large company usually does not qualify for the free tier.
- Do license conditions apply to the images, text or video a model produces?
- Sometimes. The FLUX Kommunity license asks commercial users to filter or review outputs and to add AI-generated disclosures where the law requires them, while OpenMDW places no restrictions on outputs. Separately, the law on who owns AI outputs differs from country to country.
- Are there licenses that only work in some countries?
- Yes. The ERNIE 5.5 Flash Base license and the Tencent Hunyuan Community License, which HunyuanOCR uses, are valid only outside the European Union, the United Kingdom and South Korea. The model page and the license page show the exact territory, so check them before you deploy in those regions.
- Can I fine-tune a model and sell the result?
- Only if the license allows derivative models, and the conditions usually follow your derivative. Apache 2.0 and MIT allow it with light conditions. Other licenses may require a credit line, keep revenue limits in place or keep the derivative under the same agreement.
Key terms in this guide
- Fine-tuning Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain.
- Gated model A gated model is an open model whose download on Hugging Face requires you to log in and accept its terms, or request access, first.
- Open weights An open-weights model is an AI model whose trained parameters, called weights, are published so that anyone can download and run it, usually under a license that sets the terms of use.
- OpenRAIL (Responsible AI License) OpenRAIL is a family of licenses that let you use, modify and share a model, including commercially, but require every copy and derivative to carry a list of banned uses.
- Permissive license A permissive license, such as Apache 2.0 or MIT, lets you use, modify and sell a model with very few conditions, usually just keeping a notice.
- Revenue cap (revenue threshold) A revenue cap is a license condition that makes a model free below a certain income and requires a separate agreement above it.
- Territory restriction A territory restriction limits where a license is valid: using the model in the excluded countries or regions is not licensed.
Models mentioned in this guide
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| Mistral Large 4 Mistral AI | 1.05T 52B active | 1M | Mistral Open-Weight (Conditional) | 671 GB GPU server | |
| Muse Spark Open Meta | 816B 42B active | 1M | Muse Model License (Conditional) | 563 GB GPU server | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU | |
| ERNIE 5.5 Flash Base Baidu | 280B 14B active | 1M | ERNIE 5.5 Flash Base (Conditional) | 193 GB Big-memory desktop | |
| Solar Open 2 Upstage | 250B 15B active | 1M | Upstage Solar (Conditional) | 157 GB Big-memory desktop | |
| HunyuanOCR Tencent | 1.12B | 128K | Tencent Hunyuan (Conditional) | 0.773 GB Laptop or phone | |
| Chandra OCR 2 Datalab | 5.3B | 256K | OpenRAIL-M (Datalab) (Conditional) | 3.66 GB Laptop or phone | |
| Supertonic 3 Supertone | 99M | – | OpenRAIL-M (Conditional) | 0.46 GB Laptop or phone | |
| Parakeet TDT 0.6B v3 NVIDIA | 627M | – | CC BY 4.0 (Permissive) | 2.89 GB Laptop or phone | |
| Pocket TTS Kyutai | 18M – 100M | – | CC BY 4.0 (Permissive) | 0.127 GB Laptop or phone |